PeSTo: parameter-free geometric deep learning for accurate prediction of protein interacting interfaces

Abstract

Predicting the interactions that a protein can establish with other molecules from its structure remains a major challenge. As shown by recent applications to tertiary structure prediction and opposite to current mainstream methods for interaction interface prediction, low-level, geometry-based, physicochemical-agnostic representations of structures have several advantages over methods that require pre-calculation of surfaces, charges, hydrophobicity, and other kinds of parameterizations. Here we introduce a new geometric transformer that acts directly on protein atoms labelled with nothing more than element names. The resulting model outperforms the state of the art for the prediction of protein-protein interaction interfaces and distinguishes interfaces with nucleic acids, lipids, small molecules and ions with high confidence. The low computational cost of this method (available online at https://pesto.epfl.ch/) enables processing high volumes of structural data, such as molecular dynamics trajectories allowing the discovery of interfaces that remain inconspicuous in static experimentally solved structures.